Slashdot Mirror
Can Anyone Suggest a Good Switch?
wgadmin asks: "I am a sysadmin for a 500-node Linux cluster. We are doubling the size of our compute nodes, so, as a consequence, we need a new switch. We currently have a Foundry FastIron 1500 -- however, a) it doesn't have enough ports (only 208) and b) it is extremely unreliable. We want something that's solid as a rock. We process mostly serial jobs. And we will probably require ~320 ports. What's everyone using for their HPC clusters? There's so many high performance switches on the market, we hardly know where to start."
What level of interconnect do you want? (gig copper? gig fiber? 10/100?)
Or are you looking for something more specialized (HIPPI compliant or something similarly obscure?)
That said, if you're looking for in the ethernet space, we've been really happy with our recent Extreme Networks chassie's. Their black diamond 10k line is the newest release, and it looks awesome. It's really dense, they've got crazy levels of backplane bandwidth, and ours have been really reliable (granted, we have the previous generation of the gear). The chassies have blades (just like everyone else) that can speak 10/100, 10/100/1000 copper, gig fiber, 10 gig fiber, etc.
Zapman
We've been using 3com switches and theyre rock solid. I was rooting for cisco a while ago because I'm studying for some certs, but the price difference is huge.
3Com comes with stackable switches, upto 8 of 48 ports which should be enough. The stacking bus is something like 10gbit or 32gbit for all-gigabit switches.
The switch market is really (1) Cisco (2) 3Com and (3) HP for market share, and I recommend you go with these. Cisco is more than 100% expensive than anyone else for the same stuff, and I've never been impressed with HP, so look at 3Com. Take a look at all those confusing nortel switches too, the number 4 of the market. You'll most likely find your switch between 3com and cisco, unless you want to give up reliability.
"Give orange me give eat orange me eat orange give me eat orange give me you." -Nim Chimpsky
Extreme Networks has a great line of switches.
The Black Diamond 10808 would work great for the type of envrionment you have setup from the sounds of it. Also, Extreme is usually 20-40% cheaper then Cisco and Foundry for the equivilant appliance.
We currently use an Alpine 3808 with 192 100mbps ports and it's never had a problem with uptime and configuration is a simple and straightforward.
Check out Mon and Mon.cgi
http://www.force10networks.com/ claim to have the higest port density.
Of course if cost is an object - I guess you are stuck on simple GbE, rather than a faster interconnect. You should look at 270 for high density interconnect... Throw in some 360's for outside connectivity and you are set.
I have mod points and I am not afraid to use them
Cisco is the de-facto brand of networking gear for standard stuff like Ethernet. How much better are these high-performance switches people are talking about as suggestions in the comments here? This is not a rhetorical question, I just realy want to know and I'm too lazy and uninterested to look into it myself, but not lazy enough to stop typing this slashdot post. Is it enough to be worth going with non-Cisco for HPC clusters that use Ethernet-based interconnects? I know Cisco isn't infallible, but for all kinds of reasons they're a good bet in networking gear come purchasing time, at least outside this HPC cluster business.
11*43+456^2
Flat neighborhood networks, basically you get to use "cheap" cards and switches in a web configuration to provide a fast interconnect between nodes.
Other than that, a Cisco 6513 with 11 10/100/1000 48 port switch cards would fit the bill to provide a single chasis switch for all 500 nodes. Hope you've got a decent budget, because it will cost you.
09f911029d74e35bd84156c5635688c0
Give serious thought to FC-IP and director-class fibrechannel kit. Performance-wise it'll thrash Ethernet, and there're various clever tricks you can do with directors clustered together via Open Trunking meaning that a bunch of 160 port boxes (a McData 6140 is your best bet here) will do as well as a larger single box.
I'm a sysadmin for a 3 large clusters in the same league, we use stackable 48 port Nortel switches. Each switch is 1u, and the interconnects don't use a separate port. The switches have wildly expensive support options, however because it just works we've never had to pay for support on them.
We use to have Foundry ourselves, but their switches were crap, they would suddenly become dumb hubs and lose their ip, etc.
We tried HP, but found their interface cumbersome and unfamiliar with weird networking related issues that would pop up.
Cisco's been rock solid, but very expensive.
Can I get an eye poke?
Dog House Forum
The fastiron is fine for the job. Upgrade to a BigIron 15k or something if you want more ports.
If your fastiron is unreliable, it's broken and you need it fixed, it's not normally broken.
The LiNX runs on BigIron 8k and Exreme Networks blackdiamonds, both of which are pretty damn good.
Send email to a few supercomputing centers. These places have tons of clusters, with lots of vendors throwing hardware at them. They're also often associated with schools, so they're not competitors and they actually -want- people to learn from what they've done.
To get you started:
http://www.ncne.org
http://www.psc.edu
http://www.sdsc.edu
http://www.ncsa.edu
Yeah, it's Pittsburgh-centric. Guess where I'm posting from. There's probably somewhere closer to you.
The things you want to figure out before calling:
-What's your budget? (Nice stuff tends to be more expensive)
-How much does latency matter? (Usually, lots. Sometimes, not so much. Put numbers here.)
-What's your architecture (at several levels of technical detail)? Can you use 64-bit PCI? Do you have to work with a proprietary bus? Can you use full-height, full-length cards? What OS -exactly- are you using? (Hint: "Linux" ain't close enough.) What version and vendor of PVM/MPI/whatever are you using, and can you switch?
--
The 6513 is basically the same thing but with four extra slots.
The 6509 chassis lists at $9.5K and the 6513 $15.25K. That's completely bare bones. The supervisor modules run anywhere from $6K to $28K at list. The 48 port 10/100/1000 modules list at $7.5K while a 24 port SFP fiber blade lists for $15K. You'll need two power supplies at $2K-5K each.
On the cheap end, to get the port density you're looking for out of Cisco, you'll pay about $70K list. But if you find the right reseller, you can see a discount of 30-40%.
All numbers in this post should be considered best guess, based on quotes I've gotten. They may be out of date. They are not official prices from Cisco. Take with the appropriate grain of salt.
Still, with a plan, you only get the best you can imagine. I'd always hoped for something better than that. -CP
I'd just say go for the most that you can afford.
:D
Our 3Coms have served us well, and between them, they work with anything from 10M ethernet to 2gig fibre optic.
I worked with a major brewer for a while, and their Cisco kit was very reliable, but it never had to handle much of a load. It did survive being kicked about, dropped, my boss's driving it several hundred miles unsecured in the back of a van. I doubt out 3com kit would have survived that!
Basically, if you can afford Cisco, go for it. If not, use 3Com.
Incidentally, if you want your server room to look cool, go for Black Diamond
They are definitly the best choice.
"Emancipate yourself from mental slavery, none but ourselves can free our minds !"
They're already running Linux, you insensitive clod! RTFAS
MacBook Pro. Worst name since the Bicycle
We're using the unmanaged HP procurve modular 1Gbps switches in our clusters, but they run VERY HOT when utilized (our switches get hammered 24/7 - like most clusters probably do) and we had some overheating issues with them. Our clusters aren't as large as yours, but I'd suggest going with a major manufacturer (IBM, HP, Cisco) if you're putting all of your eggs in one basket (switch-wise).
One thing is get a switch that's modular (most good ones are), but if something goes out, you'll only loose 8 or 32 nodes instead of the whole switch.
Nortel's Passport 8600 384 ports per chassis, true wire-speed, redundant everything, layer 2-7 switching. Also, if you need more ports simply add another 8600 and use Multi-link-trunking (MLT) between the switches. Wash Rinse Repeat. Networks that use these are smokin!
Of course, if you are looking for the typical Ask Slashdot for free solutions answer you can forget it. These puppies cost a bundle.
Whoa! "RTFAS"... dang, it's a good thing the OP ducked.
try a hickory tree. Stings like hell and the mere thought is a deterrant for most rascals and rapscallions.
I can't really give you a solution at the moment... ;o)
but are you hiring?
...Rob
The American Dream isn't an SUV and a house in the suburbs; it's Don't Tread On Me.
isn't almost all the latency in your network from software? why not build a hypertree from cheap 24-port switches? At $200 a pop, you could make a
tree with 12 roots for $8000. spend more to get
more cross-section bandwidth, less for less. it
scales with your budget.
-I like my women like I like my tea: green-
We made the mistake of buying a bunch of 3COM switches a while back because they were so much cheaper. We got what we paid for too. We're currently having to replace them all because every piece is failing now that the warranties are expired. They were just built really cheap-o. Cooling fans are dying, electrolytic caps in power supplies failing. A few switches just suddenly started continually rebooting themselves for no reason. If you only need them to work for the duration of the warranty and intend to replace them immediately afterwards, then 3COM's are a cheap buy. Cisco's last forever, even long after their technology is obsolete, their hardware is built like army tanks... but you do pay for it too.
Go with Cisco Catalyst switches. Because of the number of systems you have on that net, I'm guessing that cost is less of an object than it would be for a small business, so go with Cisco. The server farm for the company I'm consulting for right now needs to be up 24/7/365, so that's what we use. They 3550 series has a bunch of different options (10/100 or 1000 and 10/100 with gigabit uplinks).
Selecting a switch fabric calls for more information than that. That your jobs are primarily serial implies that latency isn't a high priority, and bisection bandwidth might even be of minimal importance but other factors come into play.
For example, are you using NFS over the network? How large are your data sats? Do you tend to just queue jobs and have them start/finish whenever, or do you tend to launch a lot of jobs at once? Do you have to transfer a data set with each job, or is it more a matter of running with different parameters over the same data set?
In some cases, simple connectivity may be all you actually need, in others, anything short of wire speed switching across the cluster will kill you.
When you run parallel jobs, do they run on the whole cluster, or just portions of it (this determines if you want to consider clos networking, or if small wire speed switches with reletively thin uplinks will be OK).